UnitY: Two-pass Direct Speech-to-speech Translation with Discrete Units
This addresses the problem of efficient and accurate speech-to-speech translation for applications requiring real-time processing, though it is incremental as it builds on existing direct translation methods.
The paper tackles direct speech-to-speech translation by proposing UnitY, a two-pass architecture that generates textual representations and then discrete acoustic units, achieving a 2.5-4.2 ASR-BLEU improvement over a single-pass model with a 2.83x decoding speed-up.
Direct speech-to-speech translation (S2ST), in which all components can be optimized jointly, is advantageous over cascaded approaches to achieve fast inference with a simplified pipeline. We present a novel two-pass direct S2ST architecture, UnitY, which first generates textual representations and predicts discrete acoustic units subsequently. We enhance the model performance by subword prediction in the first-pass decoder, advanced two-pass decoder architecture design and search strategy, and better training regularization. To leverage large amounts of unlabeled text data, we pre-train the first-pass text decoder based on the self-supervised denoising auto-encoding task. Experimental evaluations on benchmark datasets at various data scales demonstrate that UnitY outperforms a single-pass speech-to-unit translation model by 2.5-4.2 ASR-BLEU with 2.83x decoding speed-up. We show that the proposed methods boost the performance even when predicting spectrogram in the second pass. However, predicting discrete units achieves 2.51x decoding speed-up compared to that case.